Re: [PATCH] virtio pmem: fix async flush ordering

[Jeff Moyer]

Dan Williams <dan.j.williams@xxxxxxxxx> writes:

> On Fri, Nov 22, 2019 at 8:09 AM Jeff Moyer <jmoyer@xxxxxxxxxx> wrote:
>>
>> Dan Williams <dan.j.williams@xxxxxxxxx> writes:
>>
>> > On Wed, Nov 20, 2019 at 9:26 AM Jeff Moyer <jmoyer@xxxxxxxxxx> wrote:
>> >>
>> >> Pankaj Gupta <pagupta@xxxxxxxxxx> writes:
>> >>
>> >> >  Remove logic to create child bio in the async flush function which
>> >> >  causes child bio to get executed after parent bio 'pmem_make_request'
>> >> >  completes. This resulted in wrong ordering of REQ_PREFLUSH with the
>> >> >  data write request.
>> >> >
>> >> >  Instead we are performing flush from the parent bio to maintain the
>> >> >  correct order. Also, returning from function 'pmem_make_request' if
>> >> >  REQ_PREFLUSH returns an error.
>> >> >
>> >> > Reported-by: Jeff Moyer <jmoyer@xxxxxxxxxx>
>> >> > Signed-off-by: Pankaj Gupta <pagupta@xxxxxxxxxx>
>> >>
>> >> There's a slight change in behavior for the error path in the
>> >> virtio_pmem driver.  Previously, all errors from virtio_pmem_flush were
>> >> converted to -EIO.  Now, they are reported as-is.  I think this is
>> >> actually an improvement.
>> >>
>> >> I'll also note that the current behavior can result in data corruption,
>> >> so this should be tagged for stable.
>> >
>> > I added that and was about to push this out, but what about the fact
>> > that now the guest will synchronously wait for flushing to occur. The
>> > goal of the child bio was to allow that to be an I/O wait with
>> > overlapping I/O, or at least not blocking the submission thread. Does
>> > the block layer synchronously wait for PREFLUSH requests?
>>
>> You *have* to wait for the preflush to complete before issuing the data
>> write.  See the "Explicit cache flushes" section in
>> Documentation/block/writeback_cache_control.rst.
>
> I'm not debating the ordering, or that the current implementation is
> obviously broken. I'm questioning whether the bio tagged with PREFLUSH
> is a barrier for future I/Os. My reading is that it is only a gate for
> past writes, and it can be queued. I.e. along the lines of
> md_flush_request().

Sorry, I misunderstood your question.

For a write bio with REQ_PREFLUSH set, the PREFLUSH has to be done
before the data attached to the bio is written.  That preflush is not an
I/O barrier.  In other words, for unrelated I/O (any other bio in the
system), it does not impart any specific ordering requirements.  Upper
layers are expected to wait for any related I/O completions before
issuing a flush request.

So yes, you can queue the bio to a worker thread and return to the
caller.  In fact, this is what I had originally suggested to Pankaj.

Cheers,
Jeff